Feishu MCP Manual

Feishu (Lark) MCP

This is the packaged action='manual' entry point and the progressive disclosure router. A normal schema-sufficient call does not need to load it first. Load one route only when the call is unfamiliar, consequential, or recovery-sensitive.

PUBLIC TOOL FAMILY: strict LTP-v2

Every call is {action, input, reasoning, summarize?}. action, input, and string reasoning are required; summarize, if present, is boolean. The root and each action input are closed: unknown fields, cross-action fields, the old flat/_reasoning form, and missing required fields fail before provider I/O.

send/reply content can be tagged text, markdown, post, complete schema-2.0 card, media, shares, or sticker. edit accepts text/markdown/post or complete card replacement, not media. See the message reference for exact unions and media source shapes.

A safe inbound start without a pending notification is feishu(action="check", input={}, reasoning="inspect incoming chats"), then read the affected chat to obtain its exact ID before replying. When a notification already carries a compound message id and untruncated text (see reference/message-semantics.md), reply or react with that id directly instead of a routine re-check/re-read.

BEFORE A SIDE EFFECT

Configuration belongs to the orchestrator/admin; an avatar must not configure or reconfigure this MCP. Setup, credential changes and rollout require explicit owner authorization; routine message schemas do not grant it.

• receive_id_type defaults to open_id; use chat_id for a group. An omitted account selects the first configured account. Verify the account and recipient before send; verify the target before reply, edit, delete, or react.
• IDs are {account_alias}:{chat_id}:{feishu_message_id}. Pass an inbound compound ID back to reply verbatim. A reply target that is gone fails; it never becomes a fresh send. Group/topic inbound messages require an explicit Bot mention (@all alone is ignored); allowed_users gates senders, and saving a contact does not grant admission.
• Each physical outbound chunk is attempted once. A partial result lists exact delivered IDs and the failed chunk; do not replay the whole action or silently downgrade it. Reconcile provider state before another lifecycle operation.

CARDS ARE THREE DIFFERENT THINGS

1. A complete schema-2.0 business card is outbound content. An authorized click becomes one deduplicated card_action inbox record and wakes the Agent; update its source_message_ref or send a fresh response. A callback record is not a reply target.
2. Local command cards (/help, /status, /kanban, /system, /refresh, /sleep, /clear, /taskcard) update themselves. Their callbacks stay local: no business inbox record and no Agent wake. Group commands still need admission and a Bot mention.
3. Automatic/programmable resident Task Cards are mechanical projections, not messages to manage with these actions. placeholder=true is separate native progress feedback; edit it only at meaningful phase changes and send the final durable answer separately.

SETTINGS SHOW

settings accepts {} only and returns a read-only inventory. Its seven row owners are linked below; SHOW never writes or grants configuration authority.

Setting config path

See reference/message-semantics.md#setting-config-path.

Setting account aliases

See reference/message-semantics.md#setting-account-aliases.

Setting account app ids

See reference/message-semantics.md#setting-account-app-ids.

Setting account app secrets

See reference/message-semantics.md#setting-account-app-secrets.

Setting account allowed users

See reference/message-semantics.md#setting-account-allowed-users.

Setting task card enabled

See reference/message-semantics.md#setting-task-card-enabled.

Setting task card normal rows

See reference/message-semantics.md#setting-task-card-normal-rows.

ONE OWNER FOR DEPTH

For failures, preserve error_code, retryable, and retry_after_seconds. max_attempts=1 makes retryable guidance a new caller decision, not an adapter retry. Never expose secrets, raw envelopes, provider keys, or local paths in external evidence.